Anomaly Detection and Attribution Using Bayesian Networks
A. Kirk, Jonathan Legg, EDWIN D. EL-MAHASSNI · 2014
Abstract : We present a novel approach to anomaly detection in Bayesian networks, enabling both the detection and explanation of anomalous cases in a dataset. By exploiting the structure of a Bayesian network, our algorithm is able to efficiently search for local maxima of data conflict between closely related variables. Benchmark tests using data simulated from complex Bayesian networks show that our approach provides a significant improvement over techniques that search for anomalies using the entire network, rather than its subsets. We conclude with demonstrations of the unique explanatory power of our approach in determining the observation(s) responsible for an anomaly.